CNS*2026 Halifax

Introduction
Dementia involves progressive cognitive decline, with Alzheimer’s disease (AD) and frontotemporal dementia (FTD) as major subtypes. Electroencephalography (EEG) provides a noninvasive and accessible measure of brain dynamics and able to capture pathological slowing, often reflected by increased theta-to-alpha power ratio (TAR) [1]. Reduced signal complexity has also been reported in dementia. We tested whether Extrema-Segmented Entropy (ExSEnt), which separates temporal and amplitude irregularity, improves discrimination and interpretability beyond conventional electroencephalography biomarkers [2].

Methods
We analyzed resting-state EEG from 88 subjects, including 36 with AD, 23 with FTD, and 29 controls [3]. After preprocessing and independent component analysis, we classified brain sources into four clusters: right and left prefrontal cortices (R/LPFC) and right and left visual associations (R/LVA). Then we computed a set of classical spectral and complexity features. We also measured the ExSEnt metrics, which quantifies the entropy of extrema-based segment durations, amplitudes, and their pair. To find table features for dementia detection, we performed stability selection with elastic-net logistic regression and nested leave-one-subject-out validation [4].



Results
ExSEnt improved discrimination mainly in the LPFC, where balanced accuracy increased from 71.5% to 82.6%. In that region, ExSEnt features showed high selection stability and replaced conventional measures as dominant predictors in 3 out of 4 brain areas. The most informative variables were temporal and joint temporal-amplitude entropy measures in alpha and theta bands, with negative coefficients indicating reduced irregularity and reduced dynamical variability in dementia. Other regions showed weaker or inconsistent gains, suggesting a localized effect rather than a global one.

Discussion
ExSEnt provides a compact and interpretable measure for EEG-based dementia classification. The results suggest that dementia is associated not only with spectral slowing but also with reduced diversity of extrema timing and temporal-amplitude variability in LPFC dynamics. Because ExSEnt is explainable by design, when combined with stability selection and sparse linear classification, it yields robust biomarkers with direct physiological interpretation. These findings support ExSEnt as a promising candidate for explainable dementia screening and motivate future validation against cognitive severity measures.



References
[1] M. Penttil¨a, J. V. Partanen, H. Soininen, P. Riekkinen, Quantitative analysis of occipital EEG in different stages of Alzheimer’s disease, Electroencephalography and Clinical Neurophysiology 60 (1985) 1–6. doi:10.1016/0013-4694(85)90942-3
[2] S. Kamali, F. Baroni, P. Varona, Exsent: Extrema-segmented entropy analysis of time series, arXiv (2025). doi:10.48550/arXiv.2509.07751.
[3] A. Miltiadous, et al., A dataset of scalp EEG recordings of Alzheimer’s disease, frontotemporal dementia and healthy subjects from routine EEG, Data 8 (2023) 95. doi:10.3390/data8060095.
[4] N. Meinshausen, P. B¨uhlmann, Stability selection, Journal of the Royal Statistical Society Series B: Statistical Methodology 72 (2010) 417–473. doi:10.1111/j.1467-9868.2010.00740.x.

Acknowledgement
This work was supported by grants PID2024-155923NB-I00 and CPP2023-010818.